Adaptive learning for image classification

Abstract

A method, a computerized apparatus and a computer program product for adaptive learning for image classification. The method comprises applying a set of classification models on a calibration dataset and on a production dataset, and calculating disagreement measurements over the predictions thereof on each dataset. Based on similarity measurement between the disagreement measurement of the calibration and the production datasets, being below a predetermined threshold, a data drift is indicated in the production dataset. The method further comprises determining a training dataset for training a classification model for the production dataset. The training dataset is selected over a plurality of sets of images ordered according to time intervals in which images therein are obtained. The selection is performed based on weights determined for the plurality of sets.

What is claimed is: 
     
         1 . A method comprising:
 imaging a product to generate an image with a camera or optical sensor, wherein the product is a flat panel display or a printed circuit board;   identifying a defect in the image;   obtaining a set of classification models, each of which is a neural network trained to predict a label for the image, wherein the predicted label indicates a class of the defect in the image, wherein each classification model of the set of classification models is configured to predict a label from a same set of labels;   applying a predictor on a production dataset, wherein the predictor is a neural network trained to predict, for the image, a label from the same set of labels;   applying the set of classification models on a calibration dataset of images thereby providing for each image of the calibration dataset of images an array of predicted labels and providing a set of arrays of predicted labels for the calibration dataset, wherein the predicted labels for the calibration dataset of images are provided by one of the classification models;   calculating a disagreement measurement of the set of classification models on the calibration dataset, wherein the disagreement measurement is calculated based on the set of arrays of predicted labels for the calibration dataset, wherein the disagreement measurement is affected by differences between predictions of classification models of the set of classification models;   applying the set of classification models on the production dataset of images thereby providing a set of arrays of predicted labels for the production dataset;   calculating a production disagreement measurement of the set of classification models on the production dataset;   determining a similarity measurement between the production disagreement measurement and the disagreement measurement;   in response to the similarity measurement being below a predetermined threshold, indicating a data drift in the production dataset; and   in response to said indicating the drift in the production dataset, retraining the predictor based on at least a portion of the production dataset.   
     
     
         2 . The method of  claim 1 , wherein the set of classification models excludes the predictor. 
     
     
         3 . The method of  claim 1 , wherein the set of classification models comprises the predictor. 
     
     
         4 . The method of  claim 1 , further comprising re-computing the disagreement measurement based on the at least a portion of the production dataset. 
     
     
         5 . The method of  claim 1 , wherein the disagreement measurement is computed based on a portion of images for which a tuple of classification models of the set of classification models provide different labels. 
     
     
         6 . The method of  claim 1 , wherein the disagreement measurement is computed based on a portion of images for which a tuple of classification models of the set of classification models provide a same label. 
     
     
         7 . The method of  claim 1 , wherein the set of classification models comprise n classification models denoted as cls 1 , . . . , cls n , wherein the disagreement measurement is computed based on a portion of images for which the set of classification models provide an array of predicted labels having a value of (l 1  . . . , l n ), where l i  is a label from the same set of labels that is predicted by cls i  wherein the value of (l 1  . . . , l n ) is a heterogeneous value. 
     
     
         8 . A computer program product comprising a non-transitory computer readable storage medium retaining program instructions, the computer program product configured to carry out the method of  claim 1 . 
     
     
         9 . A computerized apparatus having a processor, the processor being adapted to perform the steps of:
 sending instructions to image a product to generate an image with a camera or optical sensor, wherein the product is a flat panel display or a printed circuit board;   identifying a defect in the image;   obtaining a set of classification models, each of which is a neural network trained to predict a label for the image, wherein the predicted label indicates a class of the defect of the image, wherein each classification model of the set of classification models is configured to predict a label from a same set of labels;   applying the set of classification models on a calibration dataset of images thereby providing for each image of the calibration dataset of images an array of predicted labels and providing a set of arrays of predicted labels for the calibration dataset, wherein the predicted labels for the calibration dataset of images are provided by one of the classification models;   applying a predictor on a production dataset, wherein the predictor is a neural network trained to predict, for the image, a label from the same set of labels;   calculating a disagreement measurement of the set of classification models on the calibration dataset, wherein the disagreement measurement is calculated based on the set of arrays of predicted labels for the calibration dataset, wherein the disagreement measurement is affected by differences between predictions of classification models of the set of classification models;   applying the set of classification models on the production dataset of images thereby providing a set of arrays of predicted labels for the production dataset;   calculating a production disagreement measurement of the set of classification models on the production dataset;   determining a similarity measurement between the production disagreement measurement and the disagreement measurement;   in response to the similarity measurement being below a predetermined threshold, indicating a data drift in the production dataset; and   in response to said indicating the drift in the production dataset, retraining the predictor based on at least a portion of the production dataset.